LED spectrum fingerprint feature individual identification method and device based on linear attention network

Abstract

The invention belongs to the field of visible light communication, and relates to an LED spectrum fingerprint feature individual identification method and device based on a linear attention network. Comprising the following steps: 1, acquiring original spectral data of an LED light source, and forming a spectral fingerprint database; 2, preprocessing original spectral data of the LED light source, constructing a training data set, training by using a SpecLinNet network, inputting the training data set into the trained SpecLinNet network, extracting feature vectors before a Softmax layer, and forming a feature vector library; and step 3, performing weighted multiple scatter correction preprocessing on to-be-identified LED original spectrum data obtained from the spectrum collector, inputting the data into the SpecLinNet network, performing cosine similarity comparison on feature vectors output by the network and features in the feature vector library one by one, and performing judgment. The method provided by the invention is a security authentication method of a physical layer, and is higher in security.

Description

Technical Field

[0001] The present invention belongs to the field of visible light communication and relates to a method and device for individual identification of LED spectrum fingerprint characteristics based on a linear attention network. Background Art

[0002] Traditional radio communications such as 5G / WiFi are common means of interconnecting everything, and play a fundamental "foundation" role in the construction of an integrated network in the air, space, land, and sea. However, constrained by factors such as electromagnetic leakage, radio communications cannot meet the application requirements of many vertical subdivided industries for both high speed and security. For example, radio communications are prohibited in some sensitive areas such as industrial workshops, nuclear power plants, and security conference rooms. Visible Light Communication (VLC) has opened up new spectrum resources, with the characteristics of ubiquitous coverage, ultra-wideband, and inherent security. It is irreplaceable in the field of medium and short-distance, high-speed, and high-density wireless interconnection.

[0003] Visible light communication uses light emitting diodes (LEDs) as light sources for both lighting and communication. Analogous to mobile communications, LED lights in visible light communication play the role of cell base stations. Therefore, untrusted LED lights (equivalent to fake base stations) that have not been authenticated pose a great challenge to the security of visible light communication systems. How to identify unauthenticated devices is an important issue in the security design of visible light communication.

[0004] Traditional visible light communication systems perform security authentication through protocols at the signal level. This authentication method does not require additional hardware equipment and has a low implementation cost, but there is a problem of counterfeiting. For high-security communication needs, this solution has risks at the theoretical level.

[0005] Device fingerprint technology generates a unique device signature by analyzing hardware characteristics and has been widely used in the field of radio frequency communications to prevent spoofing or counterfeiting attacks. The physical basis of this solution is the inconsistency of the device during the manufacturing process, that is, each device has its own unique subtle fingerprint characteristics, which can be used as the basis for individual identification. The spectral fingerprint and electrical characteristics of the LED can both reflect the unique properties of the device and can be used for individual identification of the device. Among them, the spectral fingerprint is a non-contact method and has more advantages in system design. At the same time, the spectral fingerprint of the LED is its original characteristic and has the outstanding advantage of not being imitated or tampered with. Summary of the invention

[0006] In the existing visible light communication system, the secure and reliable authentication of LED light sources is achieved through the communication protocol at the electrical signal level, which is likely to be counterfeited and tampered with, and cannot fundamentally solve the problem of fake base stations. In order to solve this problem, the present invention provides an individual identification method and device of LED spectral fingerprint characteristics based on a linear attention network, which is based on the spectral fingerprint characteristics of the LED light source itself. The fingerprint characteristics are caused by the inevitable inconsistency in the production process, are the original characteristics of the LED device, are unique, and cannot be tampered with or counterfeited. It is a physical layer security authentication method with higher security.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] In a first aspect, the present invention provides a method for individual identification of LED spectrum fingerprint characteristics based on a linear attention network, comprising:

[0009] Step 1: Collect the original spectrum data of the LED light source and form a spectrum fingerprint database;

[0010] The original spectrum data of the LED light source is collected by the receiver using a spectrum collector, and the spectrum collector is specifically a portable spectrometer. The spectrum range includes the entire range of the LED radiation spectrum, the spectrum range includes 350-750nm, and the measurement step is designed to be between 0.1-10nm according to actual needs. Spectral collection is carried out under various conditions such as the relative distance between the transmitting and receiving ends, the relative angle, and the noise level.

[0011] Step 2: Preprocess the original spectral data of the LED light source in the spectral fingerprint feature database to construct a training data set, and use the SpecLinNet network for training, then input the training data set into the trained SpecLinNet network, extract the feature vectors before the output layer, and form a feature vector library. In the SpecLinNet network model, the first layer is a fully connected layer, the second layer is a linear attention layer, followed by a fully connected layer and an output layer; the preprocessing refers to weighted multivariate scattering correction of the original LED spectral data to minimize the spectral scattering effect without losing the integrity of the physical characteristics.

[0012] Step 3: The original spectral data of the LED light source to be identified is first preprocessed with weighted multivariate scattering correction and input into the SpecLinNet network. The feature vector output by the network is compared with the features in the feature vector library obtained in step 2 by cosine similarity one by one: if the feature similarity with a certain LED in the feature vector library exceeds the threshold, the spectrum is determined to be derived from a known LED light source device, and the LED light source device number is output; if the feature similarity with all LED light sources in the feature vector library does not exceed the threshold, the spectrum is determined to be derived from an unknown LED light source device, that is, an untrusted device.

[0013] The spectral data preprocessing method comprises:

[0014] Step 1: Calculate the global mean spectrum:

[0015] Define the spectral matrix X∈R n×p , where n is the number of samples, p is the number of wavelengths (characteristic dimension) of each spectrum, and the spectral vector X of the i-th sample is defined as i , then use formula (1) to calculate the mean μ of the spectral vector of this type:

[0016]

[0017] Step 2: Calculate the Euclidean distance between the sample spectrum and the global mean spectrum:

[0018] For each sample X i , its Euclidean distance D from the mean spectrum μ i It is defined as formula (2):

[0019]

[0020] Step 3: Calculate the weighted reference spectrum:

[0021] First, use formula (3) to calculate the weight W of the current sample i :

[0022]

[0023] Among them, λ is the regularization parameter;

[0024] Then, the weighted reference spectrum R is calculated using formula (4): i :

[0025] R i =W i X i +(1-W i )μ (4);

[0026] Step 4: Correct the spectrum using the least squares method:

[0027] Using formula (5), for each sample X i Perform a least squares fit:

[0028] X i =aR i +b (5);

[0029] Among them, coefficients a and b are obtained by linear regression fitting respectively;

[0030] Step 5: Multiple Scatter Correction (MSC):

[0031] The original spectrum is normalized using formula (6) to obtain the corrected spectrum

[0032]

[0033] The above preprocessing method can effectively reduce the spectral scattering effect while retaining key physical characteristics, thereby more accurately reflecting the true characteristic information of the sample.

[0034] In the SpecLinNet network model of the present invention, the first layer is a fully connected layer that receives the original spectral input, the second layer is a linear attention layer, followed by a fully connected layer and an output layer, wherein the output layer is a Softmax layer for outputting classification probabilities. Before entering the linear attention layer, the feature h relu First convert to a two-dimensional tensor x′∈R 1×256 , and then processed using a linear attention model. relu It is a commonly used activation function in neural networks.

[0035] The SpecLinNet network of the present invention can accurately capture the key characteristic areas of specific spectral bands in the LED spectrum, reduce the interference of redundant information, and thus improve the model's ability to extract features from the LED spectrum.

[0036] The linear attention model processing includes:

[0037] Assume that the input x is a three-dimensional tensor, represented by x∈R N×T×d , where: N represents the batch size. d represents the dimensionality of each input vector. The input x is mapped to the weight matrix W corresponding to Query, Key, and Value through three different linear transformations. Q ∈R d×d , W K ∈R d×d , W V ∈Rd×d ;

[0038] These weight matrices project the input vector into different subspaces via equation (7):

[0039] Q = xW Q , where Q∈R N×T×d ,

[0040] K=xW K , where K∈R N×T×d ,

[0041] V=xW V , where V∈R N×T×d (7);

[0042] The attention score matrix A is calculated using formula (8):

[0043]

[0044] Where d is the dimension of the Query and Key vectors.

[0045] Apply the Softmax function to all scores of each Query to convert the attention scores into weights. The Softmax function is as follows:

[0046]

[0047] A ij represents the correlation between position i and position j in the sequence;

[0048] The attention weight α is calculated as follows:

[0049]

[0050] Among them, α∈R N×T×T ;

[0051] The attention output matrix O is calculated using formula (11):

[0052] O = αV(11);

[0053] Each row of the attention output corresponds to an element of the original input sequence, which is obtained by weighted summing of all other elements.

[0054] The present invention also provides an LED spectrum fingerprint feature individual identification device based on a linear attention network, comprising:

[0055] LED light sources, including trusted LED light sources and untrusted LED light sources;

[0056] The server controls the LED light source at the transmitting end to radiate a light signal through a driving circuit, wherein the light signal includes a modulated communication signal; the untrusted LED light source is driven by the malicious server to construct a fake base station;

[0057] The receiver identifies the individual LED light source through its spectrum information, and the receiver includes a spectrum collector, a PD photoelectric detector and a receiving processing module.

[0058] The beneficial effects of the present invention are:

[0059] The spectral fingerprint feature used by the method of the present invention is based on the original characteristics of the LED light source and is related to the physical properties of the device itself. It has the outstanding advantage of being tamper-proof and non-counterfeitable, thereby achieving physical layer security.

[0060] The present invention can be applied to visible light communication, visible light indoor positioning, fixed asset registration and other applications.

[0061] (1): In a visible light communication system, any LED light source device (such as an LED lamp) can be individually identified through the spectral fingerprint information of the LED light source device before accessing the communication. If the LED light source device is determined to be unauthorized (untrusted device, similar to a pseudo base station in mobile communications), the LED light source device is prohibited from accessing the communication network. This is a contactless method that solves the security authentication problem in visible light communication.

[0062] (2): Indoor positioning in shopping malls, basements, etc. After the LED light source equipment is installed, each LED light source equipment has a fixed and known position, so indoor positioning can be completed by individually identifying the LED light source equipment.

[0063] (3): In the fixed asset registration and management of government agencies, enterprises and institutions, for any equipment with LED light sources such as operating status indicator lights, hard disk indicator lights, keyboard lights, etc., the LED spectrum fingerprint individual recognition method can be used to achieve unique and tamper-proof identification of fixed assets. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a schematic diagram of the LED spectrum fingerprint feature individual identification system proposed in the present invention.

[0065] Figure 2 Flowchart of the LED spectrum fingerprint feature individual identification method based on linear attention network of the present invention.

[0066] Figure 3 Comparison of spectra before and after preprocessing: (a) original spectrum, (b) spectrum after preprocessing.

[0067] Figure 4The network modules proposed in this invention are (a) SpecLinNet network structure and (b) Linear Attention module structure. FC (fully connected) represents the fully connected layer. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution in the embodiment of the present invention will be clearly described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] Example 1

[0070] The LED light spectrum fingerprint feature individual recognition system of this embodiment is as follows: Figure 1 As shown in the figure, a trusted LED light is installed at a fixed position in the room. At the same time, there may be a malicious LED light source (untrusted LED light) in the room. Both are transmitting LED light sources (i.e., light source transmitters). The server controls the transmitting LED light source to radiate light signals through the driving circuit. The light signal itself contains the modulated communication signal. The untrusted LED light builds a pseudo base station under the drive of the malicious server. The receiver can identify the individual LED lights through the spectral information to achieve the purpose of positioning and secure communication. The hardware structure of the system includes a control server, a transmitting LED light source, a receiver, etc., wherein the receiver includes a spectrum collector, a PD photodetector, and a receiving processing module.

[0071] like Figure 2 As shown, the LED spectrum fingerprint feature individual identification method based on the linear attention network (SpecLinNet) of this embodiment includes:

[0072] Step 1: The receiver uses a spectrum collector to collect the original spectrum data of the LED lamp and form a spectrum fingerprint feature database. The spectrum collector is specifically a portable spectrometer. The spectrum range includes the entire range of the LED radiation spectrum to obtain as many subtle features as possible. Generally speaking, the spectrum range includes 350-750nm, and the measurement step can be designed to be between 0.1-10nm according to actual needs. Spectrum acquisition is carried out under various conditions such as the relative distance, relative angle, and noise level of different transceivers to maximize the generalization performance of the later recognition model. Step 2: In the training stage, the original spectrum data of the LED lamp in the spectrum fingerprint feature database is preprocessed to construct a training data set, and the SpecLinNet network is used for training to build an accurate closed set recognition model. The preprocessing refers to weighted multivariate scattering correction of the original spectrum data of the LED lamp to minimize the spectrum scattering effect without losing the integrity of the physical characteristics. Then, the training data set is input into the trained SpecLinNet network, and the feature vector before the Softmax layer is extracted to form a feature vector library.

[0073] Step 3: In the recognition stage, the original spectrum data of the LED lamp to be identified obtained from the spectrum collector is first preprocessed with weighted multivariate scattering correction and input into the SpecLinNet network. The feature vector output by the network is compared with the features in the feature vector library obtained in step 2 one by one by cosine similarity: if the feature similarity with a certain LED lamp in the feature vector library exceeds the threshold, the spectrum is determined to be derived from a known LED device, and the LED lamp number is output; if the feature similarity with all LED lamps in the feature vector library does not exceed the threshold, the spectrum is determined to be derived from an unknown LED device, that is, an untrusted device. For signals identified as unknown devices, they will be manually labeled and stored in the database (feature vector library) as new categories of data for accumulation and update. This step not only enhances the breadth of the database, but also promotes a new round of network training, thereby continuously optimizing and expanding the recognition capabilities of the model.

[0074] Scattering effects can cause distortion of spectral signals, affecting the quantitative and qualitative analysis of components. Multiple Scatter Correction (MSC) corrects scattering effects to make the sample spectrum closer to the true material characteristics. The present invention uses weighted multiple scattering correction to preprocess the raw spectral data of the LED light source, and uses the relationship between the dynamic equilibrium sample spectrum and the global mean spectrum. According to the different characteristics of the sample spectrum, different correction strategies are applied to each sample, while reducing the spectral scattering effect, retaining valuable characteristic information, and ensuring that the characteristics of different samples are properly considered.

[0075] The spectral data preprocessing method comprises the following contents:

[0076] Step 1: Calculate the global mean spectrum:

[0077] Define the spectral matrix X∈R n×p , where n is the number of samples and p is the number of wavelengths (characteristic dimension) of each spectrum. Define the spectral vector X of the i-th sample i , then the mean μ of this type of spectral vector is:

[0078]

[0079] Step 2: Calculate the Euclidean distance between the sample spectrum and the global mean spectrum:

[0080] For each sample X i , its Euclidean distance D from the mean spectrum μ i Defined as:

[0081]

[0082] Step 3: Calculate the weighted reference spectrum:

[0083] Using distance D i , calculate the weight W of the current sample i :

[0084]

[0085] Among them, λ is the regularization parameter to avoid the weight imbalance caused by too large or too small distance value. Then, the weighted reference spectrum R is calculated i , based on the weighted average of the samples:

[0086] R i =W i X i +(1-W i )μ (4)

[0087] Step 4: Least squares correction of the spectrum:

[0088] For each sample X i , we weight the reference spectrum R i Perform the least squares fitting. Specifically:

[0089] X i =aR i +b (5)

[0090] Among them, coefficients a and b are obtained by linear regression fitting.

[0091] Step 5: MSC correction:

[0092] Corrected spectrum The original spectrum is normalized to obtain:

[0093]

[0094] By calculating the distance between the sample and the global mean, we can measure the "degree of deviation" of the sample from the whole. The closer the sample spectrum is to the mean spectrum, the more typical its behavior is and does not require too much correction. Weighted reference spectrum R i It is between the sample spectrum and the global mean spectrum, providing a more representative reference and avoiding over-smoothing caused by over-reliance on the global mean. The comparison of the spectrum before and after preprocessing is shown in the figure. Figure 3 shown.

[0095] In a fully connected neural network (FCNN), each layer of "neurons" is fully connected to each neuron in the next layer. There is no connection between neurons in the same layer, and there is no direct connection between layers. Each neuron in the lower layer receives input from all neurons in the previous layer. Each eigenvalue of the input is multiplied by the corresponding weight, plus the bias term, and a new eigenvalue is output after nonlinear transformation of the activation function. Through such multi-layer feature extraction, the neural network can gradually extract useful information from the input and finally get the expected results.

[0096] The SpecLinNet network proposed in the present invention combines the linear attention mechanism with the design of a fully connected neural network, and is mainly composed of three parts: a fully connected layer, a linear attention layer, and subsequent fully connected layers and output layers. The structure of the SpecLinNet network model of the present invention is as follows: Figure 4 As shown in (a), the first layer is a fully connected layer that receives the original spectral input, the second layer is a linear attention layer, and the last layer is a Softmax layer (with a fully connected layer in front of it) that outputs the classification probability. The structure of the linear attention module is as follows: Figure 4 As shown in (b), before adding the linear attention mechanism, for a specific batch N, the feature h relu (h relu It is a commonly used activation function in neural networks) first converted into a two-dimensional tensor x′∈R 1×256(After the network input passes through the first fully connected layer, the size is N×256, 256 represents a one-dimensional feature vector. Before inputting into the linear attention, N×256 needs to be converted to N×1×256, which means converting the feature from a one-dimensional tensor to a two-dimensional tensor. R represents the dimension of the feature vector), and then processed by the linear attention module. The core of the linear attention mechanism is to enhance specific features by learning the importance of different features. Suppose the input x is a three-dimensional tensor, represented as x∈R N×T×d , where: N represents the batch size, T represents the number of channels, where T = 1; d represents the dimensionality of each input vector; the input x is mapped to the weight matrix W corresponding to Query, Key, and Value through three different linear transformations. Q ∈R d×d , W K ∈R d×d , W V ∈R d×d ;

[0097] These weight matrices project the input vector into different subspaces:

[0098] Q = xW Q , where Q∈R N×T×d ,

[0099] K=xW K , where K∈R N×T×d ,

[0100] V=xW V , where V∈R N×T×d (7)

[0101] The attention score is calculated by the dot product between the query and the key, which measures the correlation between each position in the input sequence. In order to avoid the dot product value being too large when the vector dimension is large, we usually divide the dot product result by (where d is the dimension of the Query and Key vectors) to achieve normalization. The attention score matrix A is calculated as follows:

[0102]

[0103] To convert the attention scores into weights, we apply the Softmax function to all the scores of each Query. The Softmax function is defined as follows:

[0104]

[0105] A ijrepresents the correlation between position i and position j in the sequence. The attention weight α is calculated as follows:

[0106]

[0107] Among them, α∈R N×T×T , the final attention output is the weighted sum of the Value vectors, and the weight is determined by the attention weight α calculated above. The attention output matrix O is calculated as follows:

[0108] O=αV (11)

[0109] Each row of the attention output corresponds to an element of the original input sequence, which is obtained by weighted summing of all other elements.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An individual identification method of LED spectrum fingerprint characteristics based on linear attention network, characterized in that: include: Step 1: Collect the original spectrum data of the LED light source and form a spectrum fingerprint database; Step 2: Preprocess the original spectral data of the LED light source in the spectral fingerprint feature database to construct a training data set, and use the SpecLinNet network for training. In the SpecLinNet network model, the first layer is a fully connected layer, the second layer is a linear attention layer, and the subsequent layers are a fully connected layer and an output layer; then input the training data set into the trained SpecLinNet network, extract the feature vector before the output layer, and form a feature vector library; Step 3: The original spectrum data of the LED to be identified is first preprocessed with weighted multivariate scattering correction and input into the SpecLinNet network. The feature vector output by the network is compared with the features in the feature vector library obtained in step 2 by cosine similarity one by one: if the feature similarity with a certain LED light source in the feature vector library exceeds the threshold, the spectrum is determined to be derived from a known LED light source, and the number of the LED light source is output; if the feature similarity with all LED light sources in the feature vector library does not exceed the threshold, the spectrum is determined to be derived from an unknown LED light source.

2. The method according to claim 1, characterized in that In step 1, the original spectrum data of the LED light source is collected by a receiver using a spectrum collector, and the spectrum collector is a portable spectrometer.

3. The method according to claim 1, characterized in that In step 1, the spectral range includes the entire range of LED radiation spectrum, the spectral range includes 350-750nm, and the measurement step is 0.1-10nm.

4. The method according to claim 1, characterized in that In step 2, the preprocessing refers to performing weighted multivariate scattering correction on the original spectral data of the LED light source.

5. The method according to claim 1, characterized in that The pretreatment method comprises: Step 1: Calculate the global mean spectrum: Define the spectral matrix X∈R n×p , where n is the number of samples and p is the number of wavelengths per spectrum, defines the spectral vector X of the i-th sample i , then use formula (1) to calculate the mean μ of the spectral vector of this type: Step 2: Calculate the Euclidean distance between the sample spectrum and the global mean spectrum: For each sample X i , its Euclidean distance D from the mean spectrum μ i It is defined as formula (2): Step 3: Calculate the weighted reference spectrum: First, use formula (3) to calculate the weight W of the current sample i : Among them, λ is the regularization parameter; Then, the weighted reference spectrum R is calculated using formula (4): i : R i =W i X i +(1-W i )μ (4); Step 4: Correct the spectrum using the least squares method: Using formula (5), for each sample X i Perform a least squares fit: X i =aR i +b (5); Among them, coefficients a and b are obtained by linear regression fitting respectively; Step 5: Multivariate Scattering Correction: The original spectrum is normalized using formula (6) to obtain the corrected spectrum 6. The method according to claim 1, characterized in that In step 3, in the SpecLinNet network model, the fully connected layer is used to receive the original spectral input, and the output layer is a Softmax layer for outputting classification probabilities.

7. The method according to claim 6, characterized in that Before entering the linear attention layer, the feature h relu First convert to a two-dimensional tensor x′∈R 1×256 , and then a linear attention model is used for processing, the h relu is the activation function in the neural network.

8. The method according to claim 7, characterized in that The linear attention model processing includes: Assume that the input x is a three-dimensional tensor, represented by x∈R N×T×d , where: N represents the batch size, d represents the dimension of each input vector; the input x is mapped to the weight matrix W corresponding to Query, Key, and Value through three different linear transformations Q ∈R d×d , W K ∈R d×d , W V ∈R d×d ; The weight matrix projects the input vector into different subspaces through equation (7): Q = xW Q , where Q∈R N×T×d , K=xW K , where K∈R N×T×d , V=xW V , where V∈R N×T×d (7); The attention score matrix A is calculated using formula (8): Where d is the dimension of the Query and Key vectors.

9. The method according to claim 8, characterized in that The output classification probability includes: Apply the Softmax function to all scores of each Query to convert the attention scores into weights. The Softmax function is as follows: A ij represents the correlation between position i and position j in the sequence; The attention weight α is calculated as follows: Among them, α∈R N×T×T ; The attention output matrix O is calculated using formula (11): O = αV (11); Each row of the attention output corresponds to an element of the original input sequence, which is obtained by weighted summing of all other elements.

10. LED spectrum fingerprint feature individual recognition device based on linear attention network, characterized in that: include: LED light sources, including trusted LED light sources and untrusted LED light sources; The server controls the LED light source at the transmitting end to radiate a light signal through a driving circuit, wherein the light signal includes a modulated communication signal; the untrusted LED light source is driven by the malicious server to construct a fake base station; The receiver identifies the individual LED light source through its spectrum information, and the receiver includes a spectrum collector, a PD photoelectric detector and a receiving processing module.

Images